摘要
Biomaterials possess extraordinary properties due to intricate structures on the microscale. Learning from these microstructures is critical for the design of high-performance materials with multiple functions. However, explicit modeling of the microstructures is not always feasible. This study developed a deep generative network with a self-attention mechanism to generate three-dimensional (3D) bioinspired microstructures. The robustness of the model was first checked by generating a series of gyroids, a mathematically well-defined microstructure, which showed excellent consistency with the desired structures. The model was then applied to the microstructure of the elk antlers, which are complex and cannot be directly expressed mathematically. The results showed that the model also performs well in complex, ill-defined biological materials. The model learned the inherent patterns, generating different structures with similar geometric features. This study demonstrates the potential of using Transformer-based deep generative models that can be used to generate novel 3D microstructures from limited high-resolution X-ray micro-computed tomography data.